This article discusses the complex sociotechnical problem of fish counting and the ongoing digital transformation in this field, likely involving AI or machine learning to improve counting methods. It explores the challenges and opportunities of applying technology to this traditional practice.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
Hydropower dam operators face strict compliance requirements from the Federal Energy Regulatory Commission (FERC) to accurately count migratory fish. Computer vision systems promise to automate this process, but the regulatory environment, data quality, and complexity of human expert judgment make the 'human-in-the-loop' a key design principle.
Large hydropower dams are regulated by the U.S. Federal Energy Regulatory Commission (FERC). FERC issues licenses for construction and operation, requiring operators to submit detailed plans and studies demonstrating compliance with environmental standards. Violations can lead to sanctions, fines, or even termination of lease rights, causing operators to lose generation rights.
Hydropower dams generate electricity by storing water but block fish migration routes, affecting spawning. The Pacific Northwest is particularly concerned because hydropower is a major electricity source there, and local salmonid species are threatened or endangered. Therefore, dams must conduct regular fish passage studies, with the core dataset being fish counts—typically tallied at fish ladders.
Manual counting is performed by trained fish identification experts who work long hours in harsh environments (poor lighting, uncomfortable temperatures), recording subtle features such as species, injuries, and wild or hatchery origin. These features may only be visible for a moment as fish swim by, requiring high expertise.
Data collection methods vary: some use paper and pencil leading to transcription errors, others collect data at different granularities (hourly, daily, or monthly), and only during certain migration seasons. Different operators use different standards, making data integration difficult. These inconsistencies drive agencies to seek technological improvements.
Some agencies are beginning to explore using computer vision and machine learning to automate fish counting. Building such systems in regulated industries faces high accuracy and compliance requirements. Typical steps include: defining the problem space, collecting and labeling data, training models, designing human verification loops, and passing compliance review.
Human-in-the-loop systems aim to combine algorithmic consistency with human expert judgment. Fish biologists can correct model errors, handle edge cases, and ensure classifications align with the latest scientific understanding. This collaboration is expected to reduce errors and bias, providing more timely and accurate data.
However, early system development requires close consultation with stakeholders (regulators, biologists, operators) to clarify technical constraints and business goals. Data labeling itself requires expert involvement, forming an iterative feedback loop.
This article is based on a first-person report from The Gradient in December 2023, where the author participated as a consultant in a fish counting project. Descriptions of FERC regulatory requirements, manual counting processes, and computer vision system design are drawn from that report, without introducing external knowledge. Some inferences (e.g., system development stages) are based on general processes mentioned in the report.
The fish counting case reveals key challenges of deploying AI in highly regulated environments: technical solutions must deeply couple with human expertise and adapt to fragmented data standards and strict compliance reviews. Successful human-in-the-loop systems are not just algorithmic problems but issues of organizational coordination and process design.
Primary report
Primary source